M-Mix: Patternwise Missing Mix for filling the missing values in traffic flow data

Real-world traffic flow data often contain missing values, which can limit its usability. Although existing deep learning-based imputation methods have shown promising results by reconstructing observed values, they often overlook certain missing patterns in the dataset and perform worse on filling real missing values. This paper addresses this issue and proposes a novel masking method called Patternwise Missing Mix (M-Mix) for masked modeling-based traffic flow data imputation. M-Mix generates masks by mixing existing missing values in the target datasets to preserve the missing pattern information, thereby enhancing the performance of imputing real missing values. Additionally, a dual-objective loss function is proposed for model optimization, which predicts masked values for higher robustness and reconstructs observed values to maintain semantic correctness. Through extensive experiments on real-world datasets, M-Mix consistently demonstrates superior performance compared to other masking methods.